[Application of support vector machine in screening neurotoxic compounds from traditional Chinese medicine]

In this study, based on web database, 324 neurotoxic compounds and 234 non-neurotoxic compounds were selected as a data set for neurotoxicity discriminative model. 6 122 molecular descriptors, including charge distribution, physicochemical and geometrical descriptors,were calculated to characterize the molecular structure of neurotoxic compounds. The combination of Cfs Subset Evaluation and Best First-D1-N5 searching was used to select molecular descriptors. A discrimination model with high accuracy was built based on the support vector machine (SVM) approach. Meanwhile, the model accuracy, sensitivity and specificity were all above 80%. Besides, 30 traditional Chinese medicine compositions with neurotoxicity were set as external validation to further verify the model accuracy,with anaccuracy of 73.333%. Using the model, 13 potential neurotoxic compounds were screened from Sophorae subprostrate Radix,4 of them were verified by literatures. The results demonstrated that the discrimination model can be applied to screen neurotoxic compounds from Chinese medicinal materials.